Prompt Wrangler has been updated to support the latest models:

The cost estimates have also been updated with the cheaper 3.5 prices.

We are excited to introduce a new feature to our Prompt Wrangler tool, . This functionality allows you to partition assistant, user, and system messages when using chat style models, including and , to enhance prompt performance.

To utilize this feature, simply insert the special message break syntax: , as illustrated below:

This syntax is equivalent to the corresponding JSON structure used in the API:

We're always working to improve Prompt Wrangler, and we hope this update will help you better manage your interactions in chat-style scenarios. Be sure to explore this feature and let us know how it works for you!

For more details about the Message Separators feature, check out our documentation.

We've added two new helpers to help users perform Random Prompt Injection (RPI). RPI is a useful technique where you inject random variables into a prompt to produce more random results. You can learn more about the technique in our blog post.

Here are the two new helpers in action:

- Pick a random integer between two numbers.

- Picks a random string amongst an array of strings

You can learn more about these two new helpers in our documentation.

The way prompt versioning works has been improved dramatically! There is no longer a production version. Instead each new version is semantically versioned just like most APIs. When call the API you can optionally append in a semantic version that is pinned to a Major, Minor or Patch. This allows you the flexibility to implement as much (or as little) versioning as you desire.

Learn more about prompt versioning here!

You can now quickly jump to the documentation, leave feedback and changelog right from the dashboard. Click on the links in the bottom left.

You can now easily leverage helpers in your prompt to format arguments. As a reminder prompts use handlebars under the hood. There are four helpers available

This restricts an argument to a maximum number of tokens. This is helpful if text might be super longer and you don't want to overload the prompt.

Formats a json argument as json:

{{{json args.json}}}

Formats a json argument as yaml. This a bit more concise then json:

{{{yaml args.json}}}

Formats a json argument that is an array of objects as a markdown table. This is usually the best choice for arrays of objects:

If you want to chose the columns and their specific order you can pass in the c parameter. This accepts a string separated list of column names (object keys).

Learn more about prompt writing in the docs here